Michelle Marcelline

🚀🚀 We just published a case study about how Athina AI (YC W23) helped a $5B insurance company build its AI-powered Risk Underwriting Engine. 🎯 Lessons learned: ✅ Visibility and control of your data pipelines are crucial for building production-grade LLM features. ✅ Iteration speed significantly impacts the development and refinement of AI models. ✅ Collaborative tools can greatly enhance team productivity and efficiency. ✅ Seamless integration with existing systems ensures smoother deployment and operation. ✅ Robust evaluation methods are essential for maintaining model performance and reliability. Link to the complete case study in the comments 👇

67

15 Comments

✍ Introducing RLHF on the Indemn Platform! 📓 Reinforcement Learning from Human Feedback (RLHF) - a fancy term for a simple goal: improve AI models with human comments. Imagine AI that not only communicates in your unique voice but also continuously improves its accuracy and helpfulness through real-time feedback from your team. With RLHF, every interaction becomes an opportunity for your AI to learn and evolve, enhancing customer interactions and overall satisfaction. Our goal is to simplify AI agent development, allowing you to deploy agents across your business that perfectly align with your voice, values, and workflows. This feature release is a big step in the right direction. Curious to see how it works? Read my colleague Ganesh Iyer's article for an in-depth look. Kudos, Ganesh, for your outstanding work! https://lnkd.in/ejKcmKf8

30

It is very impressive how the basic tenets of computer/data science, set in stone in all relevant master degree and other certifications programs, just got upended: "Perhaps the most unexpected aspect of our findings is that feature learning can happen independently of improvements in both training and test loss. Note that this is hard to observe in the context of neural networks as non-zero training loss is required for the training process. Not only does this finding reinforce the narrative of rapid emergence of skills in neural networks, it is also not easily explicable within the framework of the existing generalization theory." Why This Challenges Existing Theories: a) Decoupling of Feature Learning and Performance Metrics: This finding suggests that significant and meaningful changes can occur in a model's internal representations without being reflected in standard performance metrics. It challenges the notion that we can reliably track a model's progress or potential for generalization using only loss or accuracy measurements. b) Rethinking the Role of Loss in Learning: In this scenario, the training loss is always zero, yet learning is still occurring. This challenges the fundamental idea that reducing training loss is necessary for learning. It also questions the utility of techniques like early stopping based on validation loss, as the model could be learning important features even when the loss isn't changing. c) Hidden Progress in Machine Learning: The results suggest that there might be "hidden" progress in machine learning models that isn't captured by standard metrics. This implies that our current understanding of how to measure and quantify learning in ML models may be incomplete. d) Emergent Behaviors: The sharp transition in test performance, despite gradual feature learning, challenges our understanding of how emergent behaviors arise in machine learning models. It suggests that emergence might be a result of crossing some threshold in feature quality, rather than a sudden change in the learning process itself. e) Generalization Theory: Classical generalization theory often relates a model's ability to generalize to its performance on the training set or to measures of model complexity. This work suggests that a model can be learning to generalize even when these traditional indicators show no change.

3

Reading through the Llama 3.1 paper, and in my opinion, this is the most telling about the state of frontier AI. The two most interesting pieces about this paragraph: 1. Llamas 1, 2, and 3 share a similar architecture, and it’s confirmed to be a Transformer. No surprise there, but the thing to keep in mind is that transformers hit the scene in 2017 and arguably sooner with Dzmitry Bahdahnau’s precursor RNNSearch. 2. The performance gains are largely in terms of better data curation, which comes as no surprise to us who have worked in AI/ML for years: data quality has always been of utmost importance.

6

I love YC Demo Day season. It’s an insane concentration of driven founders building something they deeply believe in at the frontier of tech. I know it's more trendy for VCs to complain about YC. But I believe Garry Tan when he says 5-10% of companies become worth a billion dollars. Maven’s hit rate from YC is even higher: 16% have had $B exits, and we’re trending toward a 40% graduation rate from Seed to Unicorn with investments we’ve made alongside YC. Since Maven’s founding in 2013, we’ve funded 12 YC startups out of the 66 companies we’ve invested in. Here are a few: Cruise: acquired by GM for $1B Embark: $4B SPAC, acquired by Applied Intuition Chariot: acquired by Ford before Series A May Mobility: raised $300M in follow on funding, including from Toyota Carrot Fertility: raised $100M in follow on funding, including from CRV and USVP Daybreak Health: Series A from Lightspeed, Series B from USV Moment Robotics: acquired by Cruise founder Kyle Vogt’s new company, The Bot Co (Maven is an investor) Our strategy differs from other YC investors. Many funds build a YC “index fund”, investing in many startups per batch. It’s not a bad move. But we focus on concentration, meeting all the companies in our thesis area and investing in one or two. Gratefully, I think every company from the list above can point to Maven as one of their most trusted partners that made a difference: intros to customers, closing follow-on funding from top VCs, connecting key hires, and offering genuine care and advice in both the best and toughest of moments. What are we looking for in a YC investment? 1) Team - a technical product team with startup DNA and hustle that can achieve a bold vision. Brilliant, determined founders no one would bet against who are ambitious enough to build an iconic company. 2) Vision worth fighting for - what are you bringing into the world that is worth all the time, effort and money required for success? Why aren’t existing solutions good enough? Are we proud to help bring this idea to the world? 3) Massive market - because of the valuation premium YC demands, even a $B outcome may not return today’s average Seed fund. Honestly, that math sucks. In order to pay the premium, we have to believe that if we’re right, we’ll be wildly successful. 4) Consumer trend - Maven invests in tech companies addressing emerging consumer trends. Even if some investments are B2B or B2B2C, they are always grounded in a consumer insight: autonomous vehicles in 2014, fertility care in 2017, mental health in 2020. Today, we’re looking at consumer applications of AI and frontier consumer health. Many successful companies we fund are outside of YC as well – like Zoom, Hello Heart, Epic!, Class, Wildtype. So we match YC’s Unicorn % across the rest of our portfolio, too 💪 YC founders in consumer tech and digital health: if you want to build an iconic company that delivers a positive impact for millions of consumers - I can’t wait to talk to you!

79

6 Comments

I'm checking out semantic search (like what you'd use in a RAG pipeline), and it turns out that the most influential part of this is how the input text is chunked (i.e. split into pieces). Here is a summary of some possible methods: (read more in Gregory Kamradt's popular tutorial https://lnkd.in/dXY4n8am)

2

2 Comments

The only podcast you might not have to listen to at 1.5X speed 😂 . Thanks for having me, EXPLORE84! I appreciate you sharing the latest and greatest of what we have been up to!

39

6 Comments